[EM] Strategy and Bayesian Regret

[Kristofer Munsterhjelm]

Jameson Quinn wrote:
> As you might have guessed, I'm arguing here for method 3. Kevin Venzke 
> has done work in this direction, but his assumptions --- that voters 
> will look for first-order strategies in an environment of highly 
> volatile polling data --- while very useful for making a computable 
> model, are still obviously unrealistic.

I can think of another method. It's somewhat unrealistic, but the amount 
of strategy used (and the kind of strategy used) would follow from the 
method itself.

In this method, just consider the election a multiplayer game, and use 
some sort of game AI to model strategy. If you want to use a multiplayer 
generalization of minmax, you'd have to turn the simultaneous game 
turn-based, e.g. the voters go in a random order and can see the votes 
of all those who came before them.

The arbitrary nature of this games-playing approach would be in how you 
model which voters are likely to strategize and how good they are at it 
(perhaps modeled as lookahead). Also, if you consider each possible vote 
a different type of turn and use a tree search like minmax, the approach 
would get impractical very quickly for ranked votes and would probably 
not be feasible at all for rated ones.